With an estimated prevalence of 12% among adults worldwide and up to 42% in the US (1, 2), obesity is a growing epidemic, causing major public health concerns (1, 3). Risk prediction and early prevention of weight gain is key to reducing the personal and global burden of obesity and its comorbidities (4). Developing obesity across the lifespan is the result of an interaction between environmental and innate biological factors, encoded by our genomes. Twin and family studies have reported heritability estimates of obesity that range between 40 - 70% (5).

In the past 15 years, genome-wide association studies (GWAS) have identified thousands of variants associated with obesity-related traits (6). Polygenic risk scores (PRSs), which are based on GWAS summary statistics, represent an individual’s overall genetic predisposition to obesity. In recent years, PRSs have been studied for their use in the prediction of future obesity and the identification of individuals at risk of obesity early on in life (7). The promise is that accurate estimation of people’s genetic predisposition would allow more targeted lifestyle intervention for those at risk. However, current PRSs, which are based on traditional GWAS, have been shown to be suboptimal, with unsolved challenges remaining (8). For example, existing methods to develop PRSs only include common variants (MAF ≥ 1%), they explain little of the variation (< 10%) in BMI and, thus, have limited ability to predict obesity (7, 9). There is a pressing need to incorporate rare variants (MAF < 1%), which have been shown to capture a proportion of the ‘missing heritability’ (10), and are currently not considered in the PRS construction.

Including rare variants in the PRS may improve the accuracy with which we estimate individuals’ genetic predisposition. Because of the large sample size of studies, such as the UK Biobank, association summary statistics for rare variants (0.1% ≤ MAF < 1%) can be assessed by single variant testing (11). However, for ultra-rare variants (MAF < 0.1%), which occur by definition very infrequently in the population, even current large-scale studies are not large enough to study their individual effects (12). The accuracy of the PRS depends largely on the power of the discovery GWAS summary statistics (13). Therefore, aggregating ultra-rare variants in genes, based on their predicted functional consequences, offers a potentially powerful complementary approach to the single variant testing (14) and subsequently, building rare variant PRSs.

The aim of our study is to leverage sequencing data from the UK Biobank and the Trans-Omics for Precision Medicine (TOPMed) program to build obesity PRSs that use rare variants (PRSs_rare) and test their associations with obesity and extreme obesity. In addition, we will test the predictive power of PRSs_rare for obesity outcomes alone or in combination PRSs_common.

In this study, we examined the contribution of rare variants to the polygenic prediction of obesity by leveraging data from 451,145 European-ancestry individuals in UK Biobank and 36,757 in TOPMed. We observed that PRSs_rare were associated with an increased risk of obesity and extreme obesity, partially independent of PRS_common. Nevertheless, their explained variance (up to 1.49%) as well as predictive accuracy were small (AUC 0.591 for obesity and 0.630 for extreme obesity), and particularly limited when considered in combination with PRS_common.

As PRSs are becoming a standard tools in translational research and clinical practice, there has been an increasing interest to study the role of rare variants, in addition to common ones, for a range of common diseases, such as breast cancer, prostate cancer, coronary artery disease (CAD) and obesity (28–31). Most previous studies that have reported on the contribution of rare variants studied the role of pathogenic variants in one or few high-penetrance genes and did not investigate their predictive accuracy at a population level (28, 29, 31). Consistent with our findings, though, these studies demonstrated that rare variants act—at least in part—independently from common variant PRSs and add to people’s polygenic susceptibility to disease (28, 29, 31). Thus, knowing an individuals’ PRS_rare, in addition to PRS_common, may contribute to identifying individuals at high risk of obesity. However, given the limited explained variance observed in our analyses, we expect that few individuals will indeed score high on both scores. Nevertheless, for these few individuals, knowing their high risk may be valuable.

Recently, a new framework was developed to aggregate rare variant burden into a rare variant PRS (30). As an example, a rare variant genetic risk score for CAD was built, using UK Biobank data. Similar to our findings for obesity and extreme obesity, a significant association of this PRS_rare with risk of CAD was observed, although the explained variation was only 0.1% of the population variance (30). We report a similar explained variance of 0.2% for obesity and 0.5% for extreme obesity. The reasons why the PRS_rare’s explained variance is small, in particular in addition to the PRS_common, are threefold. First, the PRS_rare was not completely independent from PRS_common, even after including only non-overlapping variants. It is likely that the true causal (rare) variants were tagged by common variants in LD. Second, any new (rare) variant added to the PRS increases the PRS’ uncertainty due to statistical noise associated with estimating a new weight (32). The PRS_rare might have suffered more from this, as accurately estimating weights for rare variants requires larger sample size in general. Third, rare variants, although more likely to have larger effects (12), are too rare to explain much of the obesity epidemic in the general population.

Consistent with the low variance explained, the predictive power by the PRS_rare over that of the PRS_common was limited. The improvement in AUC for obesity (from 0.708 to 0.710) was negligible, although the AUC for the PRS_rare alone was up to 0.59. This supports our observation that the predictive power of the PRS_rare in part overlapped with that of the PRS_common. So far, no other studies have reported on the contribution of PRS_rare, in the presence of PRS_common.

In addition to using BMI summary statistics to build PRSs and test their predictive performance for obesity and extreme obesity, we built PRSs_common and PRSs_rare based on obesity and extreme obesity GWAS summary statistics. The PRS_common and PRS_rare based on BMI-GWAS summary statistics outperformed those based on obesity or extreme obesity GWAS summary statistics, which is in line with previous findings that PRS_common based on the full distribution explains a larger proportion of the variance than when based on the tails of the distribution (33). For the ultra-rare variants, the PRS_rare-burden based on obesity summary statistics performed better than the those based BMI-based summary statistics, which maybe be due to the role of ultra-rare variants in (extreme) obesity, but less in BMI. Our discovery GWASs were conducted in a relatively healthy and less deprived UK Biobank population (34), which may have limited our ability to capture the genetic contribution of rare variants for obesity and extreme obesity.

We acknowledged that our samples for analyses were restricted to one ancestry only. We focused our analyses on European-ancestry populations for which the most data are available. Because allele frequencies, LD patterns, and effect sizes, differ between ancestries, the accuracy of European-derived PRSs decays rapidly when applied to other ancestries (35). PRSs derived from other ancestries are currently underpowered because of relatively small sample sizes. As more data becomes available for other ancestries, both GWAS as well as sequencing data, the here described analyses should be performed to examine whether observation are generalizable across ancestries. Furthermore, we focused solely on obesity, a common multifactorial trait that is moderately heritable. While many complex traits have similar feature, we cannot guarantee that our observations can be extrapolated to other outcomes as the genetic architecture, explained variance from common variants, and contribution from rare pathogenic variants may differ (36).

Taken together, we demonstrate that while rare variants, aggregated in PRSs_rare, have been shown to independently associate with obesity risk, they provide a minimal improvement in prediction accuracy over PRS_common in predicting obesity risk in the general population. Our findings cast an important light on the potential value of rare variants in the prediction of complex diseases, such as obesity.

Publicly available datasets were analyzed in this study. UK Biobank data can be found here: UK Biobank (https://www.ukbiobank.ac.uk/). All TOPMed data for each participating study can be accessed through dbGaP with the corresponding accession number listed in Acknowledgments.

All phenotypic and genetic data were collected with approval from the Institutional Review Board with patient consent at each institution. This study was approved by the Institutional Review Board (IRB) of the Icahn School of Medicine at Mount Sinai in New York, New York.

Study concept and design: ZW and RL. Acquisition of cohort level data: EB, RL, ZW, NC, MF, SR, BP, JAB, JCB, ES, M-LM, ER, WK, RV, C-TL, RM, LY, RRK, DA, RK, KN, AJ, SH, MA, JR, XG, LL, SSR, PE, SL, JB, MS, DD, MG, CA, DC, CK, RJ, and AR. Statistical analysis: ZW and SC. Interpretation of data: ZW, PFO, and RL. Manuscript writing group: ZW, PFO, SC, and RL. Supervision: PFO and RL. All authors contributed to the article and approved the submitted version.

The views expressed in this manuscript are those of the authors and do not necessarily represent the views of the National Heart, Lung, and Blood Institute; the National Institutes of Health; or the U.S. Department of Health and Human Services.

BP serves on the Steering Committee of the Yale Open Data Access Project funded by Johnson & Johnson. PE has received sponsored research support from Bayer AG and from IBM Research and has also served on advisory boards or consulted for Bayer AG, Quest Diagnostics, MyoKardia and Novartis. SL receives sponsored research support from Bristol Myers Squibb/Pfizer, Bayer AG, Boehringer Ingelheim, Fitbit, and IBM, and has consulted for Bristol Myers Squibb/Pfizer, Blackstone Life Sciences, and Invitae. ES has received grant support from GSK and Bayer.

The handling editor declared a past co-authorship with one of the authors RL.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.